Towards detecting influenza epidemics by analyzing Twitter messages


 Joshua Arnold
 1 years ago
 Views:
Transcription
1 Towards detecting influenza epidemics by analyzing Twitter messages Aron Culotta Southeastern Louisiana University Department of Computer Science Hammond, LA ABSTRACT Rapid response to a health epidemic is critical to reduce loss of life. Existing methods mostly rely on expensive surveys of hospitals across the country, typically with lag times of one to two weeks for influenza reporting, and even longer for less common diseases. In response, there have been several recently proposed solutions to estimate a population s health from Internet activity, most notably Google s Flu Trends service, which correlates search term frequency with influenza statistics reported by the Centers for Disease Control and Prevention (CDC). In this paper, we analyze messages posted on the microblogging site Twitter.com to determine if a similar correlation can be uncovered. We propose several methods to identify influenzarelated messages and compare a number of regression models to correlate these messages with CDC statistics. Using over 500,000 messages spanning 10 weeks, we find that our best model achieves a correlation of.78 with CDC statistics by leveraging a document classifier to identify relevant messages. Categories and Subject Descriptors H.3.3 [Information Search and Retrieval]: Text Mining General Terms Algorithms, Experimentation, Measurements Keywords data mining, regression, classification, social media 1. INTRODUCTION There has been growing interest in monitoring disease outbreaks using the Internet, typically either by mining newspaper articles mentioning flu illnesses [9, 18, 1, 23, 3, 15], mining the frequency of visits to healthrelated websites [12] or mining search engine logs for flurelated queries [6, 22, 8]. The recent emergence of microblogging services such Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. 1st Workshop on Social Media Analytics (SOMA 10), July 25, 2010, Washington, DC, USA. Copyright 2010 ACM $ as Twitter.com presents a promising new data source for Internetbased surveillance because of message volume, frequency, and public availability. Initial work in this direction includes Ritterman et al. [24], who show that Twitter messages can improve the accuracy of market forecasting models by providing early warnings of external events like the H1N1 outbreak. More recently, de Quincey and Kostova [5] have demonstrated the potential of Twitter in outbreak detection by collecting and characterizing over 135,000 messages pertaining to the H1N1 virus over a one week period. However, to our knowledge, no one has measured the correlation between patterns in Twitter messages and national health statistics. In this paper, we investigate several models to analyze Twitter messages in order to predict rates of influenzalike illnesses in a population. The U.S. Centers for Disease Control and Prevention (CDC) publishes weekly reports from the US Outpatient Influenzalike Illness Surveillance Network (ILINet). ILINet monitors over 3,000 health providers nationwide to report the proportion of patients seen that exhibit influenzalike illnesses (ILI), defined as fever (temperature of 100 F [37.8 C] or greater) and a cough and/or a sore throat in the absence of a known cause other than influenza. 1 While ILINet is a valuable tool in detecting influenza outbreaks, it suffers from a high cost and slow reporting time (typically a one to two week delay). To alleviate this burden, Ginsberg et al.[8] show that an accurate regression model of ILI rates can be estimated using the proportion of flurelated queries submitted to the Google search engine over the same time period. Central to the approach is a method to select which keywords to monitor, done by computing the correlation coefficient of each keyword on heldout data. While the model in Ginsberg et al. exhibits a compelling correlation with CDC statistics over a long time period, there are several reasons to consider a model based on Twitter messages: Full messages provide domain experts with more descriptive information than queries to characterize an outbreak. Twitter profiles often contain semistructured metadata (city, state, gender, age), enabling a more detailed statistical analysis. (Note that techniques to infer demographics from search history [27] could be applied here as well. It is worth investigating whether 1
2 the predictions will be more accurate using Twitter messages.) Ginsberg et al. note that an unusual event, such as a drug recall... could cause... a false alert. Indeed, the final queries used in their regression model are not published over concerns that users may be inclined to submit some of the queries out of curiosity, leading to erroneous future estimates of the ILI percentage. While our proposed method is certainly not immune to false alerts, we demonstrate that a simple document classification algorithm can filter out many erroneous messages. Despite the fact that Twitter appears targeted to a young demographic, it in fact has quite a diverse set of users. The majority of Twitter s nearly 10 million unique visitors in February 2009 were 35 years or older, and a nearly equal percentage of users are between ages 55 and 64 as are between 18 and All our data sources are publicly available, enabling reproducibility and followon research. We collect over 500,000 Twitter messages from a 10 week period and develop several regression models that predict ILI rates based on the frequency of messages containing certain keywords. We find the following results: 1. Aggregating keyword frequencies using separate predictor variables (i.e., multiple linear regression) outperforms aggregating keyword frequencies into a single predictor variable (i.e., simple linear regression). 2. Selecting keywords based on residual sum of squares appears to be more effective than selecting keywords based on correlation coefficient. 3. A simple bagofwords classifier trained on roughly 200 documents can effectively filter erroneous document matches, resulting in better model predictions. This final model achieves a.78 correlation with CDC statistics over 5 weeks of validation data. 2. MODELING INFLUENZA RATES Let P be the true proportion of the population exhibiting ILI symptoms. In all experiments, we assume P is the value reported by the CDC s ILINet program. Let W = {w 1... w k } be a set of k keywords, let D be a document collection, and let D W be the set of documents in D that contain at least one keyword in W. We define Q(W, D) = D W to be the fraction of documents in D that D match W. We propose estimating P from Q(W, D) using linear regression. We consider several competing systems, which vary based on the number of regression coefficients, how the keywords W are chosen, and whether matching documents D W are filtered for false matches. 2.1 Regression Models We consider two common regression models, Simple Linear Regression and Multiple Linear Regression. 2 Twitter older than it looks. Reuters MediaFile blog, March 30th, Simple Linear Regression Following Ginsberg et al.[8], we first consider a simple linear model between the logodds of P and Q(W, D): logit(p ) = β 1logit(Q(W, D)) + β 2 + ɛ (1) with coefficients β 1, β 2, error term ɛ, and logit function logit(x) = ln( X ). 1 X In this model, there is a single Q variable. We refer to this as Simple Linear Regression Multiple Linear Regression Because W contains more than one keyword, it is natural to consider expanding Simple Linear Regression to include separate parameters for each element of W. This results in the Multiple Linear Regression model: logit(p ) = β 1logit(Q({w 1}, D)) β k logit(q({w k }, D)) + β k+1 + ɛ (2) where w i W. Ginsberg et al.[8] found Multiple Regression to overfit in initial experiments, and so only report results using Simple Linear Regression. We investigate whether this holds for Twitter data as well. Both Simple and Multiple regression models are fit using ridge regression with regularization parameter of 1.0E Keyword Selection We now turn to the problem of selecting W, the keywords likely to be contained in messages reporting ILI symptoms. We consider two methods of generating W, which will determine the value of Q(W, D) used in each of the regression models Selection by Correlation Coefficient We adopt the method of Ginsberg et al.[8], which creates W from a larger set W by iteratively adding terms in order of their correlation with heldout validation data. For each word, we fit a Simple Linear Regression model to observed data using only that word as W. We then evaluate each model on validation points, from which we can calculate the correlation between the predicted and true values. Keywords are then added to W in order of decreasing correlation. In more detail, given a set of n ILI values P = {P 1... P n} used for training data, we estimate the correlation of each term w i W as follows: 1. Loop for j = 1 to n: (a) Reserve P j for validation, and fit a Simple Linear Regression Model on the remaining points P \ P j using W = {w i}. (b) Predict ˆP j using the learned regression model. 2. Collect all predicted ˆP j values into ˆP, and compute the Pearson correlation coefficient between ˆP and P. Note that it is possible to perform an additional crossvalidation procedure to determine when to stop adding terms to W, but in this work we simply report results over a range of sizes for W.
3 2.2.2 Selection by Residual Sum of Squares We repeat the same term selection procedure as above, but replace the Pearson correlation coefficient with the residual sum of squares (RSS): ILI RSS(P, ˆP) = X i (P i ˆP i) 2 (3) 2.3 Keyword Generation The Keyword Selection techniques in the previous section first require a larger set of keywords W from which to select. We propose two methods to generate the initial candidate list of words W Handchosen keywords We first consider a simple set of four keywords consisting of {flu, cough, sore throat, headache} 3. Besides the word flu, the additional three terms are commonly known flu symptoms Most frequent keywords To expand this candidate set, we search for all documents containing any of the handchosen keywords. We then find the top 5,000 most frequently occurring words in the resulting set. The idea is to seed the selection process with terms that are likely to correlate with ILI rates without manually tuning this set. 2.4 Document Filtering As Ginsberg et al.[8] point out, a principal drawback of using keyword matching to calculate Q is that it is susceptible to fluctuations in term frequency that are unrelated to influenza. For example, the keyword Tylenol may be a valuable term, but the recent recall of several Tylenol products led to a spike in its term frequency, without a corresponding spike in ILI rates. We propose a firststep to mitigate this problem by training a classifier to label whether a message is reporting an ILIrelated event or not. This is related to problems such as sentiment analysis [21] and textual entailment [7], which in their most general form can be quite difficult due to the ambiguities and subtleties of language. We limit this difficulty somewhat here by only considering documents that have already matched the handchosen ILIrelated terms of Section The classifier then calculates a probability that each of these messages is in fact reporting an ILI symptom. We train a bagofwords document classifier using logistic regression to predict whether a Twitter message is reporting an ILI symptom. Let y i be a binary random variable that is 1 if document d i is a positive example and is 0 otherwise. Let x i = {x ij} be a vector of observed random values, where x ij is the number of times word j appears in document i. We estimate a logistic regression model with parameters θ as: p(y i = 1 x i; θ) = e ( x i θ) (4) We learn θ using LBFGS gradient descent [16] as implemented in the MALLET machine learning toolkit 4. We use the default regularization parameter Λ = 1. 3 Note that the sore throat query is treated as a conjunction of terms, not a phrase. 4 Figure 1: Ground truth ILI rates, as obtained from the CDC s weekly tracking statistics. s 04 are used for training; weeks 59 are used for testing. Number of messages 574,643 Number of tokens 12.3 million Vocabulary size 582,419 Average number of tokens per message 21.4 Characteristics of collected Twitter mes Table 1: sages. We incorporate the classifier into a Simple Regression Model by defining the expected fraction of positively classified documents as P d Q p(w, D) = i D W p(y i = 1 x i; θ) (5) D This procedure can be understood as weighting each matched document in D W by the probability that it is a positive example according to the classifier. With an accurate classifier, we can reduce the impact of erroneous matches by limiting their impact on the computation of Q p(w, D). 3. EXPERIMENTS Below we describe the data and systems used for document filtering and ILI rate prediction. 3.1 Data We collected 574,643 Twitter messages for the 10 weeks from from February 12, 2010 to April 24, 2010 (Table 1). Note that we are restricted to messages that are set to public visibility by the poster. In order to obtain a pseudorandom sample of messages, we searched for messages containing any of the following common words: {a, I, is, my, the, to, you}. We note that this is in contrast to de Quincey and Kostova [5], who search specifically for flurelated Twitter messages. In order to regress with ILI rates, which are percentages, we need to estimate the percentage of Twitter messages that report an ILI. This requires a reliable denominator that counts all Twitter messages. Corresponding ILI rates were obtained directly from the
4 Positive Examples Headace, cold, sniffles, sore throat, sick in the tummy.. Oh joy!! : ( me too... i have a headache my nose is stopped up and it hurts to swallow :/ im dying, got flu like symptoms, had to phone in ill today coz i was in yest n was ill and was making mistakes :( Negative Examples swine flu actually has nothing to do with swin. #OMGFACT to the point where they tried to rename the virus Links between food, migraines still a mystery for headache researchers are you eating fruit breezers. those other the yummy ones haha. the other ones taste like well, cough drops haha. Table 2: Six Twitter messages labeled as positive or negative examples of an ILIrelated report. A total of 206 messages were labeled to train the classifier of Section 2.4. Accuracy F1 Precision Recall (1.9) 90.2 (1.5) 92.8 (1.8) 88.1 (2.0) Table 3: Results of 10fold cross validation on the message classification task, with standard errors in parentheses. CDC website 5. The true ILI values are displayed in Figure 1, with a horizontal line indicating the split between training and testing points. To generate labeled data for the classification model of Section 2.4, we searched for Twitter messages containing any of the four handchosen keywords defined in Section 2.3.1, making sure the messages were posted outside of the date range of the previously collected messages. (The classification messages were collected from April 25 to May 4.) This resulted in 206 messages, which were manually categorized into 160 positive examples and 46 negative examples. (Examples are shown in Table 2.) 3.2 Systems Table 4 displays the 10 different systems we evaluate. For the systems that use keyword selection, we indicate how many words it selects in parentheses. For example simplehandrss(2) selects two keywords to compute Q. It is important to note that the classification methods (classificationhand, classificationfreq) use the entire set of respective keywords to identify candidate messages, which are then assigned confidence values by the classifier to calculate Q P (W, D) as in Equation RESULTS Below we report both the accuracy of the document filter as well as the correlation and residuals of the regression models. 4.1 Document Filtering We first evaluate the accuracy of the message classifier. We perform 10fold cross validation on the 206 labeled messages, obtaining a mean accuracy of 84.29% (Table 3). Inspecting the learned parameters of the model reveals large positive weights assigned to words like headache (0.88), n (0.45) (a common abbreviation for and), have (0.40), throat (0.29), and sick (0.24). Note that because all the labeled examples already match one of {flu, cough, sore throat, 5 headache}, the classifier is learning only to classify this subset of documents. It would not be expected to perform well on messages in general. This explains why common words like n and have receive high weights, as in cough n sneezin and I have a headache. 4.2 Regression Table 5 compares the results of each of the 10 systems. Note that only the best performing keyword selection systems are displayed in this table. More detailed results are displayed in subsequent figures. We make the following four observations of the results: (1) Multiple Regression outperforms Simple Regression. Contrary to results reported in Ginsberg et al.[8], simple linear regression performs much worse than multiple linear regression. In fact, only one simple linear regression model produces a positive correlation with the testing data, and that model uses only a single keyword (flu). Adding a disjunction of additional terms only reduces the effectiveness of the simple regression model. A possible explanation for why the reverse was found for the query data in Ginsberg et al. is that queries exhibit less ambiguity (e.g., flu symptoms, influenza symptoms ). In contrast, individual keywords can be highly ambiguous (e.g., cough, sick ) and can have frequent idiomatic uses (e.g., cough it up, I m sick of politics ). Combining these disparate queries into one fraction Q(W, D) likely results in a noisy estimate of true ILI reports. Multiple regression can alleviate this problem by placing separate weights on each keyword. For example, system multihandrss(2) selects keywords flu and sore throat with coefficients 498 and 190, respectively. Despite the fact that multiple regression outperforms simple regression, multiple regression does indeed begins to overfit when too many keywords are added. In Figure 8, we can see a dramatic drop in correlation after five keywords are added. This is not unexpected, as the number of parameters for the multiple regression model (6) is now greater than the number of training points (5). (2) Keyword selection is prone to overfitting. The best performing keyword selection system that selected from the large set of 5,000 terms is multifreqrss(3), which achieves a correlation of.582. However, inspecting this system reveals that it selects the terms net, rely, leave, which do not have any obvious connection to influenza. The term net occurs most frequently with hyperlinks and leave occurs most frequently in phrases like I have to leave. Examining the training correlation and RSS reveals that these models that can choose from such a large set of keywords often over
5 Regression Classification Keyword Generation Keyword Selection classificationhand Simple yes handchosen no classificationfreq Simple yes mostfrequent no simplehandrss Simple no handchosen RSS simplehandcorr Simple no handchosen correlation simplefreqrss Simple no mostfrequent RSS simplefreqcorr Simple no mostfrequent correlation multihandrss Multiple no handchosen RSS multihandcorr Multiple no handchosen correlation multifreqrss Multiple no mostfrequent RSS multifreqcorr Multiple no mostfrequent correlation Table 4: The 10 different systems evaluated. Each system varies by the type of regression model, whether it uses classification, the method to generate candidate keywords, and the method of selecting keywords from that candidate set. Train Test system r RSS ˆσ r RSS ˆσ classificationhand e7 4.4e e classificationfreq e7 5.5e e simplehandrss(1) e e e simplefreqrss(8) e e e simplehandcorr(4) e e e simplefreqcorr(8) e e e multihandrss(2) e e e multifreqrss(3) e e e multihandcorr(1) e e e multifreqcorr(3) e e e Table 5: Pearson s regression coefficient (r), residual sum of squares (RSS), and standard regression error (ˆσ) of each system. For each of the keyword selection systems, we have displayed the system with the highest correlation on the testing data, with the number of keywords chosen shown in parentheses. Additional results by number of keywords are displayed in Figures 69. fit they obtain the highest training correlation and lowest training RSS. To obtain more reliable results from keyword selection, it is likely necessary to obtain a greater number of training points. (3) Classification appears to be an effective method of removing erroneous messages. The system achieving the highest correlation on the testing data is classificationhand (.78). Comparing this result to that of simplehandrss(1) (.498) suggests that using the classifier to compute Q results in a more reliable indicator of ILI rates. Inspecting the classifier predictions on the unlabeled data shows that the most common types of errors involve a colloquial use of cough, as in U know that ability where you re able to sense if someone is engaged in a conversation or not? Some people dont have it. cough mymom cough. While the system multihandrss(2) is competitive with classificationhand (.739 vs..78 correlation), it is important to note that multihandrss is very sensitive to the number of keywords selected. The other results are multihandrss(1) =.498, multihandrss(3) =.732, multihandrss(4) =.991. On the other hand, classificationhand uses all four handchosen keywords, relying on the classifier to determine how much each message should be weighted in computing Q. (4) Classification is sensitive to the set of training messages. This result is not too surprising, but it should be noted that the classifier s performance diminishes considerably when applied to documents with a very different word distribution. System classificationfreq obtains a poor.396 correlation on the testing data because it must classify documents that match any of the 5,000 keywords created by the Most Frequent Keywords method described in Section This differs from classificationhand, in which it classifies documents that match the handchosen keywords, which is the same type of documents that were labeled to train the classifier. (Note, though, that there were no actual overlapping documents, since the labeled documents were drawn from a different time range). Finally, while the best model exhibits a high correlation with true ILI rates (.78), its residual term is still likely too high for practical use. There are at least two reasons for this. First, the 10 weeks of data only results in 10 ILI data points. More data will likely enable both a better estimate of performance as well as a reduction in overfitting observed in the keyword selection procedures. We are in the process of collecting more messages, and fortunately, this data should be available soon, as Twitter has agreed to donate to the U.S. Library of Congress every public Twitter message since its inception (March 2006) [25]. Secondly, the characteristics of the traintest split make prediction particularly difficult. As Figure 3 shows, while the five training points exhibit limited fluctuation, a sharp drop in ILI rates begins with
6 multihandrss(2) truth multifreqrss(3) truth Figure 2: Results for multihandrss(2) on training and testing data points. Figure 4: Results for multifreqrss(3) on training and testing data points classificationhand truth simpelhandrss(1) truth Figure 3: Results for classificationhand on training and testing data points. the first test point as fluseason winds down. A training sample covering an entire flu season would likely result in significant error reductions. 5. RELATED WORK There have been a number of proposals to monitor emerging health threats through an automated analysis of online news sources [18, 1, 23, 15]. While many rely on keyword matching or document classification, some apply more complex linguistic analysis such as namedentity recognition and topic modeling [9, 3]. A similar correlation was found in flurelated queries submitted to search engines [12, 6, 22, 8]. Our methodology most closely follows that of Ginsberg et al. [8]. The principal differences are the use of Twitter data, multiple regression, and document filtering. Other Web mining techniques applied to usergenerated Web content generally rely on a descriptive analysis derived by analyzing keywords and link structures. For example, Mishne et al. [19] track the mood of the blogosphere by Figure 5: Results for simplehandrss(1) on training and testing data points. tracking frequencies of terms like tired or happy, and Chau & Xu [2] analyzes link patterns to discover communities of hate groups. While this type of analysis is useful for generating descriptive trends of online data, our work aims to map these Internet trends to realworld statistics, and to accumulate these into actionable knowledge to be used by decision makers. Most other related work in analyzing usergenerated Web content is motivated by marketing applications, in which a client determines the impact of a marketing strategy by mining changes in consumer reception of the product from blog posts, product reviews, or Twitter messages [11, 14, 13]. A few researchers have applied regression on the output of a sentiment analysis system to predict real world values. For example, Gruhl et al. [10] use handcrafted keywords applied to Web documents to predict product sales. Similarly, Liu et al. [17] perform sentiment classification on blog posts, using the output to estimate sales performance of a product. They use a similar computation as our expected fraction in
7 1 1 multifreqrss simplefreqrss correlation coefficient (r) correlation coefficient (r) multihandrss simplehandrss number of keywords number of keywords Figure 6: Correlation results for simplehandrss and multihandrss on training and testing data points as a function of the number of selected keywords. Figure 8: Correlation results for simplefreqrss and multifreqrss on training and testing data points as a function of the number of selected keywords. correlation coefficient (r) multihandcorr simplehandcorr correlation coefficient (r) multifreqcorr simplefreqcorr number of keywords Figure 7: Correlation results for simplehandcorr and multihandcorr on training and testing data points as a function of the number of selected keywords. Equation 5. Additionally, de Choudhury et al. [4] show promising results predicting changes in stock market value based on properties of blogs. Most recently, O Connor et al. [20] performed perhaps the most comprehensive Twitter analysis to date, collecting one billion messages over two years, revealing that the frequency of certain handselected keywords correlates highly with public opinion polls. The results of Tumasjan et al. [26] suggest that it might be possible to use Twitter to predict election results. 6. CONCLUSIONS & FUTURE WORK In this paper, we have explored the possibility of detecting influenza outbreaks by analyzing Twitter data. It appears that different approaches are required than those used for number of keywords Figure 9: Correlation results for simplefreqcorr and multifreqcorr on training and testing data points as a function of the number of selected keywords. Note that correlation and RSS resulted in the same keywords, so there is no difference between these results and those in Figure 8. analyzing query log data. In particular, the longer messages allow us to use simple classification techniques to prevent erroneous messages from overwhelming model estimates. Additionally, it is likely that supplying the classifier with more sophisticated linguistic features (ngrams, synonyms, etc.) will improve its accuracy. Perhaps more importantly, given the informal syntax and number of spelling mistakes in Twitter messages, it is likely that a more sophisticated preprocessing stage could improve the quality of analysis. For future work we plan not only to analyze a larger number of messages, but also to associate geographic information with each message in order to perform a more finegrained regression with regionspecific ILI data reported by the CDC.
8 7. ACKNOWLEDGMENTS We thank Troy Kammerdiener for helpful discussions. This work was supported in part by a grant from the Research Competitiveness Subprogram of the Louisiana Board of Regents, under contract #LEQSF( )RDA REFERENCES [1] J. Brownstein, C. Freifeld, B. Reis, and K. Mandl. Surveillance sans frontieres: Internetbased emerging infectious disease intelligence and the HealthMap project. PLoS Medicine, 5: , [2] M. Chau and J. Xu. Mining communities and their relationships in blogs: A study of online hate groups. Int. J. Hum.Comput. Stud., 65(1):57 70, [3] N. Collier, S. Doan, A. Kawazeo, R. Goodwin, M. Conway, Y. Tateno, H.Q. Ngo, D. Dien, A. Kawtrakul, K. Takeuchi, M. Shigematsu, and K. Taniguchi. BioCaster: detecting public health rumors with a webbased text mining system. Bioinformatics, [4] M. De Choudhury, H. Sundaram, A. John, and D. D. Seligmann. Can blog communication dynamics be correlated with stock market activity? In Proceedings of the nineteenth ACM conference on Hypertext and hypermedia, pages 55 60, [5] E. de Quincey and P. Kostkova. Early warning and outbreak detection using social networking websites: the potential of twitter, electronic healthcare. In ehealth 2nd International Conference, Instanbul, Tirkey, September [6] G. Eysenbach. Infodemiology: tracking flurelated searches on the web for syndromic surveillance. In AMIA: Annual symposium proceedings, pages , [7] D. Giampiccolo, B. Magnini, I. Dagan, and B. Dolan. The third pascal recognizing textual entailment challenge. In Proceedings of the ACLPASCAL Workshop on Textual Entailment and Paraphrasing, pages 1 9, Prague, June [8] J. Ginsberg, M. H. Mohebbi, R. S. Patel, L. Brammer, M. S. Smolinski, and L. Brilliant. Detecting influenza epidemics using search engine query data. Nature, 457, February [9] R. Grishman, S. Huttunen, and R. Yangarber. Information extraction for enhanced access to disease outbreak reports. Journal of Biomedical Informatics, 35(4): , [10] D. Gruhl, R. Guha, R. Kumar, J. Novak, and A. Tomkins. The predictive power of online chatter. In Proc. 11th ACM SIGKDD Intl. Conf. on Knowledge Discovery and Data Mining, pages 78 87, New York, NY, USA, ACM. [11] M. Hu and B. Liu. Mining and summarizing customer reviews. In Proc. 10th ACM SIGKDD Intl. Conf. on Knowledge Discovery and Data Mining, [12] H. Johnson, M. Wagner, W. Hogan, W. Chapman, R. Olszewski, J. Dowling, and G. Barnas. Analysis of web access logs for surveillance of influenza. MEDINFO, pages , [13] J. Kessler and N. Nicolov. Targeting sentiment expressions through supervised ranking of linguistic configurations. In 3rd Int l AAAI Conference on Weblogs and Social Media, San Jose, CA, May [14] S. Kim and E. Hovy. Extracting opinions, opinion holders, and topics expressed in online news media text. In ACL Workshop on Sentiment and Subjectivity in Text, [15] J. Linge, R. Steinberger, T. Weber, R. Yangarber, E. van der Goot, D. Khudhairy, and N. Stilianakis. Internet surveillance systems for early alerting of health threats. Eurosurveillance, 14(13), [16] D. C. Liu and J. Nocedal. On the limited memory BFGS method for large scale optimization. Math. Programming, 45(3, (Ser. B)): , [17] Y. Liu, X. Huang, A. An, and X. Yu. ARSA: A sentimentaware model for predicting sales performance using blogs. In Proceedings of the 30th annual international ACM SIGIR conference on Research and development in information retrieval, [18] A. Mawudeku and M. Blench. Global public health intelligence network (GPHIN). In 7th Conference of the Association for Machine Translation in the Americas, [19] G. Mishne, K. Balog, M. de Rijke, and B. Ernsting. MoodViews: Tracking and searching moodannotated blog posts. In International Conference on Weblogs and Social Media, Boulder, CO, [20] B. O Connor, R. Balasubramanyan, B. R. Routledge, and N. A. Smith. From Tweets to polls: Linking text sentiment to public opinion time series. In International AAAI Conference on Weblogs and Social Media, Washington, D.C., [21] B. Pang and L. Lee. Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2(1 2):1 135, [22] P. Polgreen, Y. Chen, D. Pennock, and N. Forrest. Using internet searches for influenza surveillance. Clinical infectious diseases, 47: , [23] A. Reilly, E. Iarocci, C. Jung, D. Hartley, and N. Nelson. Indications and warning of pandemic influenza compared to seasonal influenza. Advances in Disease Surveillance, 5(190), [24] J. Ritterman, M. Osborne, and E. Klein. Using prediction markets and Twitter to predict a swine flu pandemic. In 1st International Workshop on Mining Social Media, [25] R. Stross. When history is compiled 140 characters at a time. New York Times, April [26] A. Tumasjan, T. O. Sprenger, P. G. Sandner, and I. M. Welpe. Predicting elections with Twitter: What 140 characters reveal about political sentiment. In International AAAI Conference on Weblogs and Social Media, Washington, D.C., [27] I. Weber and C. Castillo. The demographics of web search. In Proceedings of the 33th annual international ACM SIGIR conference on Research and development in information retrieval, 2010.
From Tweets to Polls: Linking Text Sentiment to Public Opinion Time Series
From Tweets to Polls: Linking Text Sentiment to Public Opinion Time Series Brendan O Connor Ramnath Balasubramanyan Bryan R. Routledge brenocon@cs.cmu.edu rbalasub@cs.cmu.edu routledge@cmu.edu Noah A.
More informationStreaming First Story Detection with application to Twitter
Streaming First Story Detection with application to Twitter Saša Petrović School of Informatics University of Edinburgh sasa.petrovic@ed.ac.uk Miles Osborne School of Informatics University of Edinburgh
More informationDiscovering Value from Community Activity on Focused Question Answering Sites: A Case Study of Stack Overflow
Discovering Value from Community Activity on Focused Question Answering Sites: A Case Study of Stack Overflow Ashton Anderson Daniel Huttenlocher Jon Kleinberg Jure Leskovec Stanford University Cornell
More informationAre Automated Debugging Techniques Actually Helping Programmers?
Are Automated Debugging Techniques Actually Helping Programmers? Chris Parnin and Alessandro Orso Georgia Institute of Technology College of Computing {chris.parnin orso}@gatech.edu ABSTRACT Debugging
More informationWhat s Strange About Recent Events (WSARE): An Algorithm for the Early Detection of Disease Outbreaks
Journal of Machine Learning Research 6 (2005) 1961 1998 Submitted 8/04; Revised 3/05; Published 12/05 What s Strange About Recent Events (WSARE): An Algorithm for the Early Detection of Disease Outbreaks
More informationTHE TRUTH ABOUT TRIPLESTORES The Top 8 Things You Need to Know When Considering a Triplestore
TABLE OF CONTENTS Introduction... 3 The Importance of Triplestores... 4 Why Triplestores... 5 The Top 8 Things You Should Know When Considering a Triplestore... 9 Inferencing... 9 Integration with Text
More informationSVMs for the Blogosphere: Blog Identification and Splog Detection
SVMs for the Blogosphere: Blog Identification and Splog Detection Pranam Kolari, Tim Finin and Anupam Joshi University of Maryland, Baltimore County Baltimore MD {kolari1, finin, joshi}@umbc.edu Abstract
More informationFake It Till You Make It: Reputation, Competition, and Yelp Review Fraud
Fake It Till You Make It: Reputation, Competition, and Yelp Review Fraud Michael Luca Harvard Business School Georgios Zervas Boston University Questrom School of Business May
More informationRecovering Semantics of Tables on the Web
Recovering Semantics of Tables on the Web Petros Venetis Alon Halevy Jayant Madhavan Marius Paşca Stanford University Google Inc. Google Inc. Google Inc. venetis@cs.stanford.edu halevy@google.com jayant@google.com
More informationMartian Chronicles: Is MARS better than Neural Networks? by Louise Francis, FCAS, MAAA
Martian Chronicles: Is MARS better than Neural Networks? by Louise Francis, FCAS, MAAA Abstract: A recently developed data mining technique, Multivariate Adaptive Regression Splines (MARS) has been hailed
More informationScalable Collaborative Filtering with Jointly Derived Neighborhood Interpolation Weights
Seventh IEEE International Conference on Data Mining Scalable Collaborative Filtering with Jointly Derived Neighborhood Interpolation Weights Robert M. Bell and Yehuda Koren AT&T Labs Research 180 Park
More informationTracking Information Epidemics in Blogspace
Tracking Information Epidemics in Blogspace Eytan Adar and Lada A. Adamic HP Labs, Information Dynamics Lab {eytan.adar,lada.adamic}@hp.com Abstract Beyond serving as online diaries, weblogs have evolved
More informationApplication of Dimensionality Reduction in Recommender System  A Case Study
Application of Dimensionality Reduction in Recommender System  A Case Study Badrul M. Sarwar, George Karypis, Joseph A. Konstan, John T. Riedl GroupLens Research Group / Army HPC Research Center Department
More informationUs and Them: A Study of Privacy Requirements Across North America, Asia, and Europe
Us and Them: A Study of Privacy Requirements Across North America, Asia, and Europe ABSTRACT Swapneel Sheth, Gail Kaiser Department of Computer Science Columbia University New York, NY, USA {swapneel,
More informationIntroduction to Data Mining and Knowledge Discovery
Introduction to Data Mining and Knowledge Discovery Third Edition by Two Crows Corporation RELATED READINGS Data Mining 99: Technology Report, Two Crows Corporation, 1999 M. Berry and G. Linoff, Data Mining
More informationPenguins in Sweaters, or Serendipitous Entity Search on Usergenerated Content
Penguins in Sweaters, or Serendipitous Entity Search on Usergenerated Content Ilaria Bordino Yahoo! Research Barcelona, Catalunya, Spain bordino@yahooinc.com Yelena Mejova Yahoo! Research Barcelona,
More informationAutomatically Detecting Vulnerable Websites Before They Turn Malicious
Automatically Detecting Vulnerable Websites Before They Turn Malicious Kyle Soska and Nicolas Christin, Carnegie Mellon University https://www.usenix.org/conference/usenixsecurity14/technicalsessions/presentation/soska
More informationPerfect For RTI. Getting the Most out of. STAR Math. Using data to inform instruction and intervention
Perfect For RTI Getting the Most out of STAR Math Using data to inform instruction and intervention The Accelerated products design, STAR Math, STAR Reading, STAR Early Literacy, Accelerated Math, Accelerated
More informationFind Me If You Can: Improving Geographical Prediction with Social and Spatial Proximity
Find Me If You Can: Improving Geographical Prediction with Social and Spatial Proximity Lars Backstrom lars@facebook.com Eric Sun esun@facebook.com 1601 S. California Ave. Palo Alto, CA 94304 Cameron Marlow
More informationUnderstanding and Combating Link Farming in the Twitter Social Network
Understanding and Combating Link Farming in the Twitter Social Network Saptarshi Ghosh IIT Kharagpur, India Naveen K. Sharma IIT Kharagpur, India Bimal Viswanath MPISWS, Germany Gautam Korlam IIT Kharagpur,
More informationYou Might Also Like: Privacy Risks of Collaborative Filtering
You Might Also Like: Privacy Risks of Collaborative Filtering Joseph A. Calandrino 1, Ann Kilzer 2, Arvind Narayanan 3, Edward W. Felten 1, and Vitaly Shmatikov 2 1 Dept. of Computer Science, Princeton
More informationIncreasingly, usergenerated product reviews serve as a valuable source of information for customers making
MANAGEMENT SCIENCE Vol. 57, No. 8, August 2011, pp. 1485 1509 issn 00251909 eissn 15265501 11 5708 1485 doi 10.1287/mnsc.1110.1370 2011 INFORMS Deriving the Pricing Power of Product Features by Mining
More informationDetecting Malware Domains at the Upper DNS Hierarchy
Abstract Detecting Malware Domains at the Upper DNS Hierarchy Manos Antonakakis *, Roberto Perdisci, Wenke Lee *, Nikolaos Vasiloglou II, and David Dagon * Damballa Inc. {manos,nvasil}@damballa.com * Georgia
More informationDefining and Testing EMR Usability: Principles and Proposed Methods of EMR Usability Evaluation and Rating
Defining and Testing EMR Usability: Principles and Proposed Methods of EMR Usability Evaluation and Rating HIMSS EHR Usability Task Force June 2009 CONTENTS EXECUTIVE SUMMARY... 1 INTRODUCTION... 2 WHAT
More informationMay All Your Wishes Come True: A Study of Wishes and How to Recognize Them
May All Your Wishes Come True: A Study of Wishes and How to Recognize Them Andrew B. Goldberg, Nathanael Fillmore, David Andrzejewski Zhiting Xu, Bryan Gibson, Xiaojin Zhu Computer Sciences Department,
More informationNo Country for Old Members: User Lifecycle and Linguistic Change in Online Communities
No Country for Old Members: User Lifecycle and Linguistic Change in Online Communities Cristian DanescuNiculescuMizil Stanford University Max Planck Institute SWS cristiand@cs.stanford.edu Robert West
More informationCan cascades be predicted?
Can cascades be predicted? Justin Cheng Stanford University jcccf@cs.stanford.edu Jon Kleinberg Cornell University leinber@cs.cornell.edu Lada A. Adamic Faceboo ladamic@fb.com Jure Lesovec Stanford University
More informationWhat people study when they study Twitter
What people study when they study Twitter Classifying Twitter related academic papers Structured Abstract Purpose Since its introduction in 2006, messages posted to the microblogging system Twitter have
More informationDetecting LargeScale System Problems by Mining Console Logs
Detecting LargeScale System Problems by Mining Console Logs Wei Xu Ling Huang Armando Fox David Patterson Michael I. Jordan EECS Department University of California at Berkeley, USA {xuw,fox,pattrsn,jordan}@cs.berkeley.edu
More informationGeneral Principles of Software Validation; Final Guidance for Industry and FDA Staff
General Principles of Software Validation; Final Guidance for Industry and FDA Staff Document issued on: January 11, 2002 This document supersedes the draft document, "General Principles of Software Validation,
More information